AI creates the first 100-billion-star Milky Way simulation - faster runs, finer physics
Date: November 16, 2025
Researchers from RIKEN, The University of Tokyo, and Universitat de Barcelona have built the first Milky Way simulation that tracks more than 100 billion individual stars. By pairing deep learning with high-resolution physics, the team removed a major bottleneck: modeling gas behavior after supernovae. The result is a model with true single-star fidelity, generated over 100 times faster than previous approaches.
Why modeling every star has been out of reach
Accurate galaxy simulations must resolve gravity, fluid dynamics, chemical enrichment, and supernova feedback across huge ranges of time and space. To capture fast events like supernova evolution, timesteps must be tiny, which explodes compute cost. That's why current state-of-the-art runs often lump roughly 100 stars into a single particle, losing small-scale behavior that matters for star formation and structure.
Even with today's best physics-first codes, fully resolving the Milky Way star by star would take about 315 hours per million years of simulated time. One billion years would stretch beyond three decades of wall-clock time, with energy demands that don't scale well by "just adding more cores."
The breakthrough: a learned surrogate for supernova feedback
The team led by Keiya Hirashima trained a deep learning surrogate on high-resolution supernova simulations. That model predicts how gas spreads during the 100,000 years after an explosion, without forcing the main simulation to step through every micro-timestep. In effect, the surrogate handles the fast physics, while the numerical solver maintains the global dynamics.
They validated the hybrid approach against large-scale runs on RIKEN's Fugaku and the University of Tokyo's Miyabi systems. Presented at SC '25, the work delivers individual-star resolution for galaxies with 100+ billion stars, at speeds previously out of reach.
Performance at a glance
- Scale: >100 billion stars, tracked individually
- Speed: 1 million years simulated in ~2.78 hours
- Throughput: ~115 days for 1 billion years (vs. >36 years)
- Detail: 100x more stars than leading earlier simulations
- Efficiency: Surrogate predicts 100k-year post-supernova gas evolution without fine timestepping
What this makes possible for astronomy
With individual-star resolution, researchers can test theories of disk structure, spiral arm dynamics, bar formation, and stellar migration with far tighter constraints. Feedback from single supernovae and local turbulence no longer vanish inside coarse particles. Chemo-dynamical histories can be compared to surveys star by star, reducing the gap between simulation and observation.
Shorter run times also enable ensembles: many realizations, parameter sweeps, and uncertainty studies that were previously impractical.
Why this matters outside astronomy
The same pattern-use a learned surrogate for the expensive, small-scale physics and a solver for large-scale flow-fits climate, weather, and ocean models. Subgrid processes (cloud microphysics, convection, mixing) are the typical bottlenecks. Replacing them with validated surrogates can bring finer grids, longer horizons, and larger ensembles within practical budgets.
How to apply the approach in your research
- Identify the bottleneck physics: Processes that force tiny timesteps or dominate runtime (e.g., feedback, microphysics, turbulence closures).
- Generate trustworthy training data: Run targeted, high-resolution patches or experiments that cover the operational regime, including edge cases.
- Constrain the model: Bake in conservation, symmetries, and stability checks. Penalize violations during training and at runtime.
- Schedule the hybrid: Decide when to call the surrogate, at what cadence, and how to hand off state to/from the solver without drift.
- Validate like you mean it: Compare against baselines on held-out scenarios, check long-horizon stability, and quantify error growth.
- Measure energy and cost: Track joules per simulated year and wall-clock time to justify scaling and ensemble sizes.
- Plan for uncertainty: Use ensembles, calibration, and diagnostics to keep confidence intervals honest.
Caveats and open questions
- Generalization: Does the surrogate remain accurate for rare events, new environments, or distributions it hasn't seen?
- Conservation and drift: Over long integrations, do small errors bias momentum, mass, or energy budgets?
- Coupling stability: How often must the full solver correct or re-anchor the surrogate to prevent artifacts?
- Data coverage: Are the training patches diverse enough to represent the parameter space researchers care about?
"I believe that integrating AI with high-performance computing marks a fundamental shift in how we tackle multi-scale, multi-physics problems across the computational sciences," says Hirashima. "This achievement also shows that AI-accelerated simulations can move beyond pattern recognition to become a genuine tool for scientific discovery - helping us trace how the elements that formed life itself emerged within our galaxy."
Sources and further reading
Keep your skills current
If you're building AI-accelerated models, a targeted upskill plan pays off. Explore practical AI courses by skill level here: Complete AI Training - Courses by Skill.
Your membership also unlocks:
